Anthropic's Mythos models remain offline as export-control talks stall
Two weeks after a U.S. export-control order forced Anthropic to suspend its Mythos-5 and Fable-5 models, negotiations with the Trump administration have made little progress, leaving the company's revenue outlook and upcoming IPO uncertain.
Anthropic has kept its Mythos-5 and Fable-5 AI models offline since a June 12 export-control order required the company to block access by any foreign national. The order, issued by the Trump administration over security concerns, applies to both non-U.S. citizens inside the United States and abroad, including Anthropic employees who are not U.S. citizens.
The company has not provided updates on the negotiations, stating that there is no new information to share. After two weeks of intensive talks, no timeline has been set for restoring the models, and it is unclear whether the order could be expanded to other firms with similar technology.
The lack of a clear regulatory framework for AI export controls has complicated the process. Unlike traditional dual-use products, which are evaluated against established checklists, Anthropic must interpret the rules on a case-by-case basis. A source familiar with the discussions said the Department of Commerce had previously tested Fable-5 without raising objections, but the issue resurfaced after a reported method for bypassing its guardrails was highlighted.
Katie Moussouris, founder and CEO of Luta Security, reviewed the reported vulnerability and argued that the issue was overstated. She wrote that allowing developers to ask the model to fix code and generate tests is a valuable security capability rather than a harmful guardrail breach.
Anthropic’s co-founder Tom Brown has taken over negotiations from CEO Dario Amodei, working alongside public-policy chief Sarah Heck, according to reports. The prolonged shutdown threatens the company’s revenue, as Mythos-class models generate higher token fees and were expected to support the firm’s upcoming initial public offering. Anthropic also faces sizable compute costs, including a reported $15 billion-per-year agreement with SpaceX for data-center access.
The uncertainty has broader implications for the U.S. AI sector. Analysts note that the government’s willingness to restrict high-performance models could affect other companies such as OpenAI, Google and Microsoft, which are developing comparable systems. Cybersecurity expert Alex Stamos warned that limiting a leading U.S. AI firm could weaken the country’s competitive position in the global AI race.
As negotiations continue, industry observers stress that any regulatory approach should balance security concerns with the need for clear, predictable guidelines for AI developers.